Integrate Knowledge Graph and Attribute Attention Mechanism to Achieve Explainable Recommendation System

  • 林 怡瑄

Student thesis: Doctoral Thesis


In the era of information explosion recommendation systems play an important role for each user in many modern Internet services Studies used ML/DL method to predict user preference in recent years However those recommendation systems have the problem of being unexplainable The problem causes human distrust of the model because humans cannot understand the mechanism within For enhancing recommendation performance recent studies unify the knowledge graph (KG) in recommended systems to get better explanations Some studies use KG as a new method to get extra features of items but they don’t solve the unexplainable problem Some studies use the existing path in KG as the input of the model to predict user preference These studies didn’t learn knowledge graph representations or have the room for improvement when the knowledge is missing in KG Another study learns KG representation and unifies KG in the recommendation However this method focuses on producing relation explanation on KG Under this circumstance when different attributes are connected to the same relation the model can’t tell which attribute is more important In our study we build the KG-based recommendation system with the attribute explanation When generating recommendation our model can show the recommended logic to illustrate which attribute is more important by the inference of KG In addition we can predict missing attributes through background knowledge and consider the missing information in recommendation To evaluate the overall performance of our model we design some experiments on separate tasks of item recommendation and attribute prediction Two different domain datasets are used to show that our method is robust on different domains The performance of our model is better than the previous recommendation models (BPR NeuMF) KG models (TransE TransH SimplE) and recommendation model unifying KG (KTUP)
Date of Award2020
Original languageEnglish
SupervisorJung-Hsien Chiang (Supervisor)

Cite this